HGS-Mapping: Online Dense Mapping Using Hybrid Gaussian Representation in Urban Scenes

TL;DR

This work tackles online dense mapping of unbounded urban scenes with limited LiDAR coverage by introducing a Hybrid Gaussian Representation that partitions the scene into sky ($G_{sky}$), road ($G_{inlier}$), and roadside ($G_{outlier}$) Gaussians. It combines LiDAR and RGB cues for initialization, uses a lightweight aircraft‑style feature matching to fill gaps, and employs a Hybrid RGBD Rasterizer with an adaptive update strategy (silhouette filter, densification, and importance pruning) to achieve online, high‑fidelity reconstruction. Empirical results on KITTI, nuScenes, Waymo, and VKITTI2 show state‑of‑the‑art rendering quality and speed, using only about $66\%$ of the Gaussians of the previous online Gaussian method. This approach enables robust, scalable urban mapping suitable for real‑time autonomous driving, with practical impact on perception and navigation in large, unbounded environments.

Abstract

Online dense mapping of urban scenes forms a fundamental cornerstone for scene understanding and navigation of autonomous vehicles. Recent advancements in mapping methods are mainly based on NeRF, whose rendering speed is too slow to meet online requirements. 3D Gaussian Splatting (3DGS), with its rendering speed hundreds of times faster than NeRF, holds greater potential in online dense mapping. However, integrating 3DGS into a street-view dense mapping framework still faces two challenges, including incomplete reconstruction due to the absence of geometric information beyond the LiDAR coverage area and extensive computation for reconstruction in large urban scenes. To this end, we propose HGS-Mapping, an online dense mapping framework in unbounded large-scale scenes. To attain complete construction, our framework introduces Hybrid Gaussian Representation, which models different parts of the entire scene using Gaussians with distinct properties. Furthermore, we employ a hybrid Gaussian initialization mechanism and an adaptive update method to achieve high-fidelity and rapid reconstruction. To the best of our knowledge, we are the first to integrate Gaussian representation into online dense mapping of urban scenes. Our approach achieves SOTA reconstruction accuracy while only employing 66% number of Gaussians, leading to 20% faster reconstruction speed.

Figures (8)

• Figure 1: Reconstruction of large scale urban scenes. Our method achieves high-quality rendering results using only two-thirds the number of Gaussians compared to the current SOTA online reconstruction method SplaTAM splatam.
• Figure 2: Overview of HGS-Mapping's pipeline. Input RGB-LiDAR stream and pose sequence are incrementally integrated into our Hybrid Gaussian Representation. We update and adjust the Gaussians with our Adaptive Update Method (Densify Control, Importance Pruning, Silhouette Filter). Our method supports rendering RGB-Depth using Hybrid Gaussian Rasterizer and exporting meshes.
• Figure 3: Effect of Gaussian Initialization & Silhouette Filter.\ref{['ablation_gaussian_init']} shows that Gaussian Initialization enhances the reconstruction quality in the upper parts of trees and buildings. \ref{['ablation_silhouette_filter']} shows that Silhouette Filter removes the abnormal depth of building's edges.
• Figure 4: Quantitative RGB results in diverse urban scenes. We compare HGS-Mapping with three off-line reconstruction methods 3DGSinstantngpmipnerf360 and an online dense mapping method splatam. The three offline methods are trained over 20K iterations but online methods only train 100 iterations per-frame. The results shows that our work achieves superior high-quality rendering while faster training speed.
• Figure 4: Ablation on Importance Pruning with different pruning rate $\eta\%$. As $\eta\%$ increases, storage size significantly declines, while there is a corresponding drop in rendering quality.